Very bad recognition with PocketSphinx

[Leo Russell]

Hello,

I need to use PocketSphinx as a speech recognition tool, and the recognition results are very bad despite the medium quality of the input files.

I try to recognise French, so we used the two Acoustic models alongside with the Language model and dictionary for French and we did not manage to get good results: only one audio file has been correctly recognised out of 288, and the other ones were not as close as the actual transcription. To do so, as I am a newbie, I only use the basic options (-hmm -lm, -dict, -infile).

Do you know if there is a way to improve the WER (which is about 85-90%), and to get better results?

I also would like to know if it was possible to improve the processing time which is a bit too long for my use.

Cheers!

[Nickolay V. Shmyrev]

To get help on the accuracy you need to describe precisely how are you using the decoder and provide the sample files you are going to recognize.

[Leo Russell]

Thank you very much for the reply and sorry for the incomplete message.

---

I tried two things :

• First, I launched PocketSphinx simply using the provided acoustic model and language model and dictionary for French > Low accuracy and very high WER. For this test, I used:
  • Both the French F0 Broadcast News Acoustic Model and the French F2 Telephone Acoustic Model acoustic models (not at a time, but I had to test them both);
  • The french3g62K.lm.dmp language model;
  • The frenchWords62K.dic dictionary
• Then, I tried to adapt the language model and to restrict the dictionary so it can fit the sample data > the WER is slightly better, but the accuracy is still very low. I only tested these changes with the French F0 Broadcast News Acoustic Model.

I both launched using this command:

pocketsphinx_continuous -dict /path/to/frenchdictionary.dic -hmm /path/to/acoustic/model/fr_FR/french -lm /path/to/language/model/frenchmodel.lm.dmp -infile /path/to/audiofile.wav

I have attached some of audio samples as well as the dictionary used for the model adaptation and the transcription file.

I hope this will help.

Thanks in advance!

[Nickolay V. Shmyrev]

Well, french model is not very good. I'll try to train Voxforge French model in a couple of days, lets see how it goes.

[Nickolay V. Shmyrev]

Well, French model has been available for some time. It's still a long way to go for the best accuracy.